International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-9 Issue-7, May 2020 Classification of Musical Instruments using SVM and KNN

S. Prabavathy, V. Rathikarani, P. Dhanalakshmi.

Abstract: Automatic classification of musical instruments is a In [6] SVM and kNN are used for the classification of challenging task. Music data classification has become a very musical instruments; MFCC, Sonogram and MFCC popular research in the digital world. Classification of the combined with Sonogram are used for feature extraction. musical instruments required a huge manual process. This Musical instruments are classified according to this method system classifies the musical instruments from a several acoustic which is used to make sound. The musical instruments were features that includes MFCC, Sonogram and MFCC combined with Sonogram. SVM and kNN are two modeling techniques classified by their material. A traditional classification was used to classify the features. In this paper, to simply musical called as eight sounds and produced by eight different kinds instruments classifications based on its features which are of materials. They are stone, metal, thread, gourd, skin, extracted from various instruments using recent algorithms. The earth, bamboo and wood. proposed work compares the performance of kNN with SVM. There are many types of musical instruments, some of the Identifying the musical instruments and computing its accuracy musical instruments are used for classification such as is performed with the help of SVM and kNN classifier, using the banjo, cello, guitar, mandolin and violin from a string combination of MFCC and Sonogram with SVM a high accuracy instrument, bass clarinet, bassoon, clarinet, contrabassoon, rate of 98% achieve in classifying musical instruments. The flute and saxophone from woodwind instrument, system tested sixteen musical instruments to find out the accuracy level using SVM and kNN. frenchhorn, trombone, trumpet and tuba from brass instrument and piano from keyboard instrument. 1284 music Keywords: Musical Instruments Classification, Mel- samples were collected from various musical instrument frequency cepstral coefficients (MFCC), Feature extraction, k- databases. The block diagram of this proposed work is Nearest Neighborhood (kNN), Sonogram, Support Vector shown is fig. 1 Machine (SVM).

I. INTRODUCTION Music is a means of communication and is an integral part of many human activities. Most ancient cultures used music for celebrations, worship, healing and preserving their stories. The fundamental characters of the tones produced are pitch, loudness, duration and timbre. One of the basic functions of a musical instrument is to produce tones of the desired pitch [1]. Classification [2] is a process of a class label assigning to the given observation. Using signal processing techniques, the raw audio signal are computed from the features in the musical analysis systems. MFCC is used to construct a feature vector. It is adequate and to Fig. 1The block diagram of Musical Instruments determine the best way to perform their transformation. This Classification. shows that MFCCs not only suitable for speech and also for music modeling [3]. SVM is used to consider a simple way II. LITERATURE REVIEW in binary classification. SVM is a conceptual learning machine that will learn from a set of training and try to For each isolated word, features were extracted and trained generalize and build correct predictions on fresh data [4]. successfully by SVM. Audio content is characterized using The performance of kNN is much better when all data the features of Sonogram in [7]. In [8] musical instruments having the same scale. It stores the training dataset which it like western were classified with HOS features were uses as its representation. combined with MFCC. The kNN classifier is used for This algorithm makes the predictions by calculating the hierarchical classification. The proposed methods in [9] like similarities between the input samples and the each training kNN, GMM, ANN and dropout ANN used to classify instance [5]. European orchestral musical instruments. The proposed algorithm [10] various types of music genres classified using the support vector machine (SVM) that can perform binary classification. In [11] the audio and video data classified into any of the five following classes such as advertisement, news, cartoon, movie and songs using Mel Revised Manuscript Received on May 07, 2020. S. Prabavathy*, Research Scholar, Department of Computer and frequency cepstral coefficients (MFCC) with colored Information Science, Department of Computer Science and Engineering, histogram features were extracted from images in video Annamalai University, Chidambaram, Tamilnadu, India. clips that used for visual features. V. Rathikarani, Assistant Professor, Department of Computer Science and Engineering, Annamalai University, Chidambaram, Tamilnadu, India. P. Dhanalakshmi, Professor, Department of Computer Science and Engineering, Annamalai University, Chidambaram, Tamilnadu, India.

Published By: Retrieval Number: G5836059720/2020©BEIESP Blue Eyes Intelligence Engineering DOI: 10.35940/ijitee.G5836.059720 1186 & Sciences Publication

Classification of Musical Instruments using SVM and KNN

In [12] speech emotion recognized using k-Nearest IV. CLASSIFIERS Neighbor (kNN) and Gaussian Mixture Model (GMM) Support vector machine (SVM) models to recognize six emotional categories such as happy, neutral, angry, surprised, sad and fearful. Emotion like Support vector machines are used for nonlinear regression ‘happy’ gives more accuracy whereas ‘fear’ and ‘surprise’ and pattern classification. It constructs the linear model, emotion gives the lowest rate. The computation speed of the with non-linear boundaries based on the support vectors to kNN classifier is fast when compared to GMM classifier. approximate the result function. The data are separated Using the Minimum Distance Classifier in [13] the musical linearly, the linear machines are trained for the optimal instruments were classified. By calculating the FRR, the hyperplane which split the data without miscalculation and results analyzed. The proposed algorithm [14] uses Auto into the most distance between the hyperplane and the Associative Neural Network (AANN) based on features nearby training points. These training points are closest to using Sonogram to recognized speech. the optimal separating hyperplane are called the support vectors. SVM architecture shows in Fig. 4.1. III. FEATURE EXTRACTION Mel-frequency cepstral coefficients (MFCC) MFCC is used to construct a feature vector. Great amount of researchers already inspecting for what purposes of MFCCs are adequate and it determines the finest way to perform those transformations. This shows that MFCCs not only suitable for speech and also for music modeling. MFCCs are widely accepted in the field of speech recognition, human speech and voice identification [15]. In [16] musical genre is classified based on the short speech signals by using MFCC as a feature extraction method. MFCC is used to characterize audio content by calculating the features. In [17] AANN is used to segment and classify the audio signal Fig. 4.1 Architecture of SVM automatically by using LPC, LPCC and MFCC as a feature extractor. The extraction of MFCC from the music signal is K-Nearest Neighbor (kNN) shown in Fig 2. kNN algorithm is very easy and very efficient. The output data is calculated while the class with the peak frequency as of the k-most related instances. For each instance the votes for their class and the class with a large amount of votes is taken as prediction [5]. Fig. 4.2 shows the process of K- Nearest Neighbor (kNN).

Fig.2 Extraction of MFCC from music signal Sonogram Sonogram is a current incarnation of feature set, music data is sampled at 22 kHz, mono format. In [7] the features of each isolated word extracted and the model is trained effectively. To characterize the audio content, Sonogram is calculated as features. SVM classifier has been used to recognize speech by learn from the training data.

Fig. 4.2 K-Nearest Neighbor

V. PROPOSED WORK In this proposed work, the music signals are preprocessed to remove noise before feature extraction, 39 MFCC features, 22 Sonogram features and 61 MFCC and Sonogram features Fig. 3 Sonogram Feature Extraction were extracted from each and every music samples. In python, a package named LibROSA which is used to In [14] spoken word converted into text using AANN model convert sound to FFT, using the audio data the MFCC & using Sonogram as acoustic feature. Feature vectors from Sonogram plot feature extraction done. In training, the saved Sonogram were given as input. VAD is used to separate features are loaded as input to the classification algorithm distinct words from the continuous speeches. The block such as SVM and kNN. Using the confusion matrix the diagram of Sonogram feature extraction is shown in fig. 3. Recall, Precision, F-Score and Accuracy were calculated.

Published By: Retrieval Number: G5836059720/2020©BEIESP Blue Eyes Intelligence Engineering DOI: 10.35940/ijitee.G5836.059720 1187 & Sciences Publication

International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-9 Issue-7, May 2020

The trained classifiers are saved as a model and inference to that model which is used to classify the class. In SVM, linear kernel is used to classify the class that one class compares the other classes like one versus rest. In kNN, the k value initialize, to classify the class iterate from class 1 to the nearest five neighbors and calculated by the highest vote. In testing, the trained features were given to the classifier algorithm as input such as kNN and SVM to classify the musical instruments. The block diagram for music instrument classification shows in fig. 5.

Fig. 6.1 (b) Classification using MFCC with kNN 6.2 Classification using Sonogram with SVM and kNN In the proposed work the features were extracted by Fig. 5 Block Diagram for Musical Instrument Sonogram and the classifiers SVM and kNN used to classify Classification the musical instruments.

VI. EXPERIMENTAL RESULTS Dataset The musical instruments sound data taken from the RWC database, musicbrainz.org, MINIM-UK musical instrument database, IRMAS: a dataset for instrument recognition in musical audio signals and NSynth dataset. The duration of music samples were range from 1 sec. to 2 min. and 1284 music samples were collected from sixteen different musical instruments from four different families such as string, woodwind, keyboard and brass. 70% of musical samples were trained and rest of data used for testing. 6.1 Classification using MFCC with SVM and kNN Fig. 6.2. (a) Classification using Sonogram with SVM In the proposed work, the features were extracted by MFCC and the classifiers SVM and kNN used to classify the musical instruments. 98% of accuracy yield by this proposed system in classifying the musical.

Fig. 6.2 (b) Classification using Sonogram with kNN

Fig. 6.1 (a) Classification using MFCC with SVM The accuracy of the proposed system kNN gives 95% and Fig. (6.1(a)) shows the classification of musical instruments 97% for SVM in classifying the musical instrument. Fig. using MFCC with SVM and fig. (6.1 (b)) shows the musical (6.2 (a)) shows the musical instrument classification using instrument classification using MFCC with kNN. Sonogram with SVM and fig. (6.2 (b)) shows the musical instrument classification using Sonogram with kNN. 6.3 Classification using MFCC and Sonogram with SVM and kNN In the proposed work the features were extracted by MFCC combined with Sonogram and the classifiers SVM and kNN used to classify the musical instruments.

Published By: Retrieval Number: G5836059720/2020©BEIESP Blue Eyes Intelligence Engineering DOI: 10.35940/ijitee.G5836.059720 1188 & Sciences Publication

Classification of Musical Instruments using SVM and KNN

SVM 88.44 81.76 84.96 97.98 Sonogram kNN 74.52 70.75 72.58 95.68

SVM 96.21 94.12 95.15 99.29 MFCC & sonogram kNN 87.11 84.78 85.92 98.17

The accuracy of SVM is good when compare to kNN in this proposed work. Fig. 7 shows the overall accuracy comparison of the proposed work.

Fig. 6.3 (a) Classification using MFCC & Sonogram with SVM 100 98 The accuracy of the proposed system kNN gives 98% and 99% for SVM in classifying the musical instrument. Fig. 96 Support Vector (6.3(a)) shows the musical instrument classification using 94 Machine MFCC and Sonogram with SVM and fig. (6.3 (b)) shows 92 k-Nearest the musical instrument classification using MFCC and Neighbor Sonogram with kNN.

Fig. 7 Accuracy comparison of proposed work

VIII. CONCLUSION In this paper, automatic classification of musical instrument system has been proposed using SVM and kNN where MFCC, Sonogram and their combination are calculated as features to classify the musical instruments. Experimental results show that, the musical data collected from various sources for classification. The combination of MFCC and Sonogram with SVM gives high accuracy of 98% and Sonogram with kNN gives the least accuracy of 95%. In this proposed work the performance of the combination of Fig. 6.2 (b) Classification using MFCC & Sonogram with MFCC and Sonogram with SVM is better when compared to kNN Sonogram with kNN. In future the musical instruments will classify using deep learning. VII. PERFORMANCE MEASURES REFERENCES Different numbers of sample set from 16 musical instruments are taken for the proposed system. The overall 1. Musical Sound, Instruments, and Equipment, By Panos Photinos performance of F-Score, precision, recall and accuracy of 2. Signal Processing Methods for Music Transcription edited by Anssi Klapuri, Manuel Davy musical instruments shows in table 1. The accuracy of every 3. Computing with Instinct: Rediscovering Artificial Intelligence by instrument is calculated using the confusion matrix. The Yang Cai musical data be trained and tested effectively; precision, 4. Learning with Support Vector Machines, By Colin Campbell, Yiming recall, F-Score and accuracy of musical instruments are Ying 5. Master Machine Learning Algorithms: Discover How They Work and Implement Them From Scratch, By Jason Brownlee 6. The Science of String Instruments, edited by Thomas D. Rossing. 7. R. Thiruvengatanadhan, Speech Recognition using Sonogram and SVM , 2018 IJRAR January 2019, Volume 06, Issue 1 www.ijrar.org (E-ISSN 2348-1269, P- ISSN 2349-5138) 8. [8] Kazi F I, Bhalke D G, Musical Instrument Classification using Table 1 Precision, Accuracy, Recall and F-Score of Higher Order Spectra and MFCC, International Conference on Musical Instruments Pervasive Computing (ICPC), -1-4799-6272-3/15/$31.00, 2015 IEEE F- 9. Miroslav MALÍK, Richard ORJEŠEK,The Comparison of Selected Precision Recall Accuracy Features Classifiers Score Audio Features and Classification Techniques in the Task of the (%) (%) (%) (%) Musical Instrument Recognition, M. Malík, R. Orješek, Musical Instrument Recognition, Poster 2016, Prague May 24. SVM 89.59 83.87 86.63 97.53 10. Lei-Ting Chen, Ming-Jen Wang, Chia-Jiu Wang, and Heng-Ming Tai, Audio Signal Classification Using Support Vector Machines, J. Wang MFCC et al. (Eds.): ISNN 2006, LNCS 3972, pp. 188 – 193, 2006. Springer- kNN 88.38 86.19 87.27 98.22 Verlag Berlin Heidelberg 2006.

Published By: Retrieval Number: G5836059720/2020©BEIESP Blue Eyes Intelligence Engineering DOI: 10.35940/ijitee.G5836.059720 1189 & Sciences Publication

International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-9 Issue-7, May 2020

11. K. Subashini, S. Palanivel, V. Ramalingam Audio-Video based Classification using SVM and AANN, International Journal of Computer Applications (0975 – 8887) Volume 44– No.6, April 2012 12. Rahul B. Lanjewar, Swarup Mathurkar, Nilesh Patel, Implementation and Comparison of Speech Emotion Recognition System using Gaussian Mixture Model (GMM) and K – Nearest Neighbor (kNN) Techniques Rahul B. Lanjewar et al. / Procedia Computer Science 49 ( 2015 ) 50 – 57 13. Sayantani Nandi, Madhura Banerjee, Parangama Sinha, Jayati Ghosh Dastidar, SVM Based Classification of Sounds from Musical Instruments using MFCC Features, Sayantani Nandi et al, International Journal of Advanced Research in Computer Science, 8 (5), May-June 2017,2144-2147 14. R. Thiruvengatanadhan, Speech Recognition Using Sonogram And AANN, International Journal of Innovative Research in Advanced Engineering (IJIRAE) ISSN: 2349-2163 Issue 01, Volume 6 (January 2019). 15. Computing with Instinct: Rediscovering Artificial Intelligence, edited by Yang Cai. 16. Gursimran Kour, Neha Mehan, Music Genre Classification using MFCC, SVM and BPNN, International Journal of Computer Applications (0975 – 8887) Volume 112 – No. 6, February 2015. 17. P. Dhanalakshmi, S. Palanivel and M. Arul, Automatic Segmentation Of Broadcast Audio Signals Using Auto Associative Neural Networks, ICTACT Journal On Communication Technology, December 2010, VOL.1, ISSUE: 04, ISSN: 2229– 6948 (ONLINE).

AUTHORS PROFILE S. Prabavathy, M.C.A,. She is currently pursuing Ph.D in Annamalai University, Chidambaram, Tamilnadu, India. Her research interests are machine learning and deep learning.

Author-1 Photo V. Rathikarani, B.E., M.E., Ph.D (CSE). Pattern Classification, Computer Networks, Medical Image Processing are her interests. She is Assistant Professor in Annamalai University, Chidambaram, Tamil Nadu, India.

P. Dhanalakshmi, B.E., M.Tech., M.B.A., Ph.D. Audio and Speech Signal Processing, Data Mining, Pattern Recognition and Software Engineering are her interests. She is Professor in Annamalai University, Chidambaram, Tamil Nadu, India.

Published By: Retrieval Number: G5836059720/2020©BEIESP Blue Eyes Intelligence Engineering DOI: 10.35940/ijitee.G5836.059720 1190 & Sciences Publication